Re: 5.3-BETA7 install cd: kernel trap 12 with interrupts disabled (fwd)

Yes, a built-in memory tester would be helpful.  However, my experience
is that many memory problems pass the x86memtest, but cause 
"seg faults" during buildworld, partly due to the extra load and heat
from disk activity.

I'm a firm believer that we should encourage our users to
buy ECC systems - motherboards that support ECC, and the
more expensive SIMMs/DIMMs that have redundant bits.

I'd rather spend a few more dollars to *know* when memory is
the problem during normal operations instead of trying to
locate the source of the problem after something bad happens.

I would be happy to lead a "documentation" project for FreeBSD
that talks about the issues and the current hardware options.
I've got the start of an article for the USENIX login magazine.
If people want to encourage me to publish, please send email
and suggest an appropriate mailing group or forum.

Thanks
-robert gray



Robert Watson <rwatson_at_freebsd.org> Fri, 15 Oct 2004 04:14:04 EDT says:
>On Fri, 15 Oct 2004, Guido van Rooij wrote:
>
>> It turns out this was a memory problem. Refitting the dimms was all
>> that was needed to 'solve' the issue. The weird thing is that
>> 1. The BIOS had memtesting enabled and did not complain
>> 2. W2K seemed to install ok (though I never let it hit the disk)
>> 3. FBSD would repeatedly crash at exactly the same spot.
>> 
>> I am beginning to wonder if we should have a boot option that enables a
>> thorough memtest from within the kernel...(e.g. boot -m). 
>
>I guess the old /dev/test_for_shoddy_workmanship driver is working :-).
>
>It's probably just bad luck -- BIOS memory testing is generally pretty
>poor, and it could just be FreeBSD stored your root vnode pointer (or some
>other critical thing) in a memory word that Win2K was using for an icon,
>so a single bit twiddle did pretty different things.  Glad it's fixed.  I
>occasionally wonder if we shouldn't build a memory tester into the FreeBSD
>boot loader to help diagnose this sort of stuff, though.  Not to run every
>boot, but as a diagnostics option.
>
>Robert N M Watson             FreeBSD Core Team, TrustedBSD Projects
>robert_at_fledge.watson.org      Principal Research Scientist, McAfee Research